VIDEO 5 ALIBABA CLOUD RESEARCH

Key Takeaways

• Global cross-border e-commerce is a $26 trillion market, and applying AI agents to it can unlock enormous value by automating complex compliance and tariff tasks.
• Current frontier AI agents fall over 50% behind super-expert levels on cross-border e-commerce search tasks, revealing a major real-world performance gap.
• The paper introduces level three theoretical rule data, requiring agents to retrieve expert knowledge and deduce a unique 10-digit HS code for each shipment.
• Agents must perform hierarchical reasoning where a single early mistake cascades into total failure, making error tolerance critical for real-world deployment.
• Rules in cross-border e-commerce have vague semantic boundaries (e.g., 'excluding') and deep logic dependencies, creating uncertainty that current agents struggle with.
• Self-improving agents that optimize without human supervision are a hot research direction, with potential to reach superhuman performance as shown by recent models.
• Memory and context management are the core bottlenecks for long-horizon agent execution, analogous to human sleep and subconscious processing for sustained performance.
• Reformulating deep search tasks as table completion offers a new perspective that significantly improves robustness and performance in long-horizon agent tasks.